Empathi is a tool for the prediction of bacteriophage protein functions. It utilizes the highly informative ProtT5
protein embeddings to make predictions. In addition, new functional groups were defined to be better suited for
machine-learning than the often-overlapping PHROG categories.
Citation:
Boulay, A., Leprince, A., Enault, F. et al. Empathi: embedding-based phage protein annotation tool by hierarchical assignment. Nat Commun 16, 9114 (2025). https://doi.org/10.1038/s41467-025-64177-5.
Getting Started
Empathi has been packaged in PyPI and as an
Apptainer container for ease of use.
The source code can also be downloaded from HuggingFace.
Prerequisites
A GPU is recommended for large datasets.
The full list of dependencies and versions can be found in requirements.txt.
Either git-lfs or Apptainer will be required. See instructions below.
Other dependencies are taken care of by pip and Apptainer.
There are three ways of installing Empathi: through PyPI, as an Apptainer container or as source code. Installation should take less than 10 minutes.
A small fasta file is provided to test installation. This should run in <1 minute.
1. PIP
First, create a virtual environment in python 3.11.5.
Download models for Empathi.
You will need git-lfs: for WSL or linux use sudo apt-get install git-lfs, for windows either use git
bash or get it from here. Then:
A fasta file of protein sequences or a csv file of protein embeddings can be used as input.
By default, a confidence >0.95 is used to assign functions. This is well suited for metagenomic datasets.
Using a high confidence threshold (--confidence 0.95) will result in more precise predictions (lower false positive rate),
but also a lower sensitivity (less proteins assigned a function). If your objective is to annotate as many proteins as possible,
consider using a confidence threshold of 0.5 (or somewhere in between), especially if working with cultured phages.
Specifying the option --only_embeddings will only compute embeddings. This step is much faster with a GPU.
The embeddings file can then be reinputted using the same command (without --only_embeddings) and specifying the new file as input file.
Options:
input_file: Path to input file containing protein sequencs (.fa*) or protein embeddings (.pkl/.csv).
name: Name of file you want to save to (wOut extension). Should be different between runs to avoid overwriting files.
--models_folder: Path to folder containing EmPATHi models. Can be left unspecified if it was added to PATH earlier.
--only_embeddings: Whether to only calculate embeddings (no functional prediction).
--output_folder: Path to the output folder. Default is ./empathi_out/.
--threads: Number of threads (default 1).
--confidence: Confidence threshold used to assign predictions (default 0.95).
--mode: Which types of proteins you want to predict. Accepted arguments are "all", "pvp", "DNA-associated", "adsorption-related", "lysis", "regulator", "cell_wall_depolymerase", "packaging", "RNA-associated", "ejection", "phosphorylation", "transferase", "nucleotide_metabolism", "reductase" and "defense_systems".
Output format
The output consists of a csv file with an annotation column regrouping all assigned annotations per protein (separated by "|") and a
column per functional category with the confidence associated to each prediction.
